The Future Arrived in 2017
Geneva. November 2017. The United Nations Convention on Certain Conventional Weapons was mid-cycle. Branka Marijan expected another dry, hypothetical slog about killer robots.
They probably wouldn’t exist. Not yet. Maybe ever.
She was wrong.
On day one, a short film called Slaughterbots played in the hall. A slick pitch deck from a fictional contractor. Drones hunting humans. Precision strikes without emotional hesitation.
“Guns don’t kill people,” the CEO said on screen. “People get emotional. Weapons don’t.”
The room shifted. Quiet. Cold.
Marijan felt the air leave the room. The scary part wasn’t the movie. It was knowing the Pentagon was building something just like it. Right then. Project Maven was already underway.
Google had just joined.
“That was the moment,” Marijan said. “The futuristic became current. We stopped talking about ‘if’ and started looking at platforms that could select targets based on sensor data alone.”
Red Lines Are Porous
Ten years later, we don’t have Terminators. Not really. But the lines have blurred.
Anthropic is the center of gravity now. A startup trying to draw red lines in blood. They ban mass surveillance. They ban autonomous target identification. Zero human involvement. No killing without a human eye on the trigger.
Anthropic is the only military AI contractor daring to set limits.
But limits are temporary things. History hates them.
Seventy years of incremental adoption. The DOD sat up in New Hampshire decades ago, seeing the potential. Now, the kill chain is faster. Faster than human thought.
Anthropic knows its fences are picket fences in a hurricane.
“It seems they know the lines won’t hold. History suggests they are already broken.”
The Hegseth Pivot
January 2026. Chaos.
Defense Secretary Pete Hegseth wanted to renegotiate every contract. He wanted to erase gray areas. “Any lawful use” became the standard. Vague. Expansive. Dangerous.
Anthropic said no.
They were the only game in town for classified network access at the time. So the Pentagon pulled a tactical masterstroke. March 2021 saw the designation: Military Supply Chain Risk.
Trump declared a ban. All government agencies told to drop Claude.
The relationship cooled. Then it warmed. Anthropic released Mythos, a cybersecurity model. The court battles raged. The silence continued. Anthropic didn’t comment.
Does it matter who won the fight?
The public finally heard the term fully autonomous weapons. It stuck.
But the work continued. The rubicon was crossed years ago. We just didn’t want to admit we were wet.
Defending by Killing?
DOD Directive 300009 sits at the heart of the mess. Written in 2012.
It says a human must be in the loop. Humans must exercise judgment.
Or do they?
Look at the Phalanx CIWS. An automated cannon on naval vessels. It shoots at missiles. Milliseconds response time. A human can’t do that. It has to.
Is that offensive?
Experts argue the distinction is semantic theater.
“One is defense if it stops a missile. Offense if it hits a bunker. But you cannot fight a war only in defense,” Sorin Adam Matei said bluntly. “It’s a choice, not a constraint.”
The 2023 update to the directive clarified nothing. The 2024 Biden memorandum still stands under Trump. But the Pentagon is gutting itself. The CDAO was restructured, isolated. Now it reports to Emil Michael. The department’s CTO.
The structure favors speed over ethics.
Stalled Talks
The CCW talks continue. Slow. Pleading.
Smaller nations want rules. Major powers want freedom.
No official definition of Lethal Autonomous Weapon exists internationally. Countries talk past each other. Some benefit from the lack of binding law.
Sarah Shoker attended these meetings. She watched decades of negotiation.
“I think everyone is tired,” she said. “Ten years. No agreement. Just words.”
From ‘Don’t Be Evil’ to Maven Smart Systems
Google broke first.
Robert Work’s memo created the Algorithmic Warfare team. AI for war zones. Google signed up.
4,000 employees revolted. Petitions. Walkouts. A moral crisis in Silicon Valley.
“We can’t outsource moral responsibility,” the staff said.
Google quit. Mid-2018.
The gap didn’t stay open.
Amazon swooped in. Microsoft followed. Palantir took over the project entirely. Project Maven became the Maven Smart System (MSS). It doesn’t just find things. It tracks them. Analyzes them. At scale.
Recent reports link MSS to the capture of Nicolás Maduro. To the strikes in Iran. Thousands dead.
Claude Enters the Chat
Anthropic got in late. But it got in.
After allowing military use in 2024, Claude entered the MSS interface.
Analysts query Claude. Ask about geography. Target types.
It’s efficient.
It makes the targeting cycle faster.
“The volume of targets makes supervision impossible by design,” Shoker said.
Claude helps. Anthropic claims the human stays in the loop. They won’t let AI press the button alone.
Good for them.
Everyone else will.
The Vacuum Fills
OpenAI signed the deal Anthropic spurned.
After snubbing Anthropic, DOD signed with eight others:
- Microsoft
- Amazon Web Services
- Nvidia
- OpenAI
- Reflection
- Oracle
- SpaceX
The market is hungry. The mission statements tweak. “Dual use” becomes “Military use.”
Shoker worked at OpenAI. She saw few people worried about decision-support risks. The community is siloed. Silent.
The companies jump at the chance. The Pentagon opens the door. The rest follows.
